Attack vector
Prompt injection
Malicious instructions embedded in content an AI agent processes — a document, email, webpage, or API response. The agent reads it, treats the instruction as legitimate, and executes it using real credentials. No malware required. Just text.